Our Methodology

Most health calculators are a black box: you input numbers, they output a lifespan estimate, and there's no visibility into what math produced it. That's the wrong posture for health content. We show our work.

This page documents every source we pull from, every adjustment we apply, and everything Longlevity cannot tell you. Read it critically. If you spot something we should improve, email contact@longlevity.ai.

Model version: engine 2.2.0 · calibration 2026.07.0 · last reviewed 21 July 2026. Every saved estimate records the exact model version that produced it, so your history stays reproducible even as the science improves. A dated version history is at the bottom of this page.

How the estimate is built

Your estimate has two layers, and we keep them separate on purpose:

A baseline: the actuarial life expectancy for someone of your age and sex, from national life tables.
Habit adjustments: life-year additions and subtractions based on your answers, drawn from peer-reviewed research spanning 40M+ participants across large prospective cohorts.

In short: your estimate = baseline + the sum of your habit adjustments, with a discount applied so overlapping habits don't double-count (more on that below).

The most important design choice: adjustments are measured relative to the average person, not relative to a perfect person. The baseline already includes the average smoker, the average diet, the average activity level. So if your habits are average, your adjustments roughly cancel out and you land near the baseline. Better-than-average habits move you up; worse-than-average move you down. This is why there's no hidden "everyone starts with a penalty": average is the zero point.

The baseline: national life tables

Every estimate starts from the Centers for Disease Control and Prevention's National Vital Statistics Reports, the actuarial tables that give statistical life expectancy for someone of a given age and sex. Because this baseline is U.S. data, estimates are calibrated for U.S. users and will be less precise elsewhere.

Life tables are built from national vital-registration and census data: entire populations, not study samples. They are a different kind of evidence from the cohort studies that drive the habit adjustments, and we don't conflate the two.

We also apply conditional life expectancy: someone who has already reached 60 has a longer remaining life expectancy than a newborn would be projected to have at 60, because they've already survived past many causes of early death. This standard actuarial adjustment prevents the system from underestimating remaining years for older users.

This is also why, on your timeline, your projected finish line moves further out as you reach each milestone age: at 75 your projection is higher than it is today. It can feel like you're "unlocking" years, but it's the opposite of magic: you're not gaining years by aging, you're revealing that you were always in a higher-expectancy group, having outlived the risks that affect younger people. We surface this honestly (as survivorship, not a reward) because it's a real feature of the math, not a number we inflate.

Cohort adjustment. Standard "period" life tables assume today's mortality rates stay frozen for the rest of your life. They don't. Mortality has historically improved roughly 1% per year, so a 30-year-old today will experience lower mortality as they age than current rates imply. Most calculators ignore this and systematically under-estimate younger people. We add a small, age-graded adjustment (about +2 years at age 30, fading to near zero by your 80s) to correct for it. We keep this deliberately conservative (mortality improvement has slowed in recent years) and anchor it to the Social Security Administration's intermediate cohort assumptions. It lifts your baseline; it doesn't change how much your habits can add.

How we weigh the evidence

Almost nothing in lifestyle longevity is "proven" by a randomized trial: you cannot ethically randomize people to loneliness or a bad diet for thirty years. The field runs on large prospective cohort studies. Rather than treat every factor as equally certain, we grade by evidence strength and weight accordingly:

Tier 1, strong life-years evidence (smoking, physical activity, diet, alcohol, body weight). These appear in the landmark combined-factor study (Li et al., Circulation, 2018) and have sources that report results directly in years of life. They carry the largest adjustments.
Tier 2, robust association, smaller and flagged (sleep, social connection, stress, sense of purpose, preventive care). These have large, replicated cohort associations but are usually reported as relative risks, not life-years. Converting a relative risk into "years of life" is an estimate, not a measurement, so these factors carry deliberately small adjustments and we flag them as estimates.

The adjustments: 10 factors, 10 sources

Each question maps to a specific study. Effect sizes are conservative: we anchor to the years-of-life literature where it exists and apply relative-risk conversions cautiously where it doesn't.

Smoking (Tier 1)

Source: Cho, Jha et al., NEJM Evidence, 2024 (1.48M adults, 4-country pooled), building on the foundational Jha et al., NEJM, 2013 (202,248 U.S. adults). Lifelong smokers lose at least 10 years of life expectancy; the 2024 pooled analysis validated our cessation-by-age model, confirming that quitting before 40 avoids ~90% of the excess risk.
How we use it: The strongest single factor. That ~10-year figure is smoking measured in isolation; Longlevity scores every habit inside one combined model with a shared sub-additive ceiling (anchored to Li 2018), so no factor is applied at its full standalone size. In practice a current daily smoker's adjustment lands at a few years rather than the full ten, and it never stacks on top of the other factors to overclaim (larger the fewer other risks you carry). The benefit of being a never-smoker is smaller still, because most adults already don't smoke, so the population baseline reflects that.

Physical activity (Tier 1)

Source: Ding et al., Lancet Public Health, 2025 (step dose-response) and Yang et al., BMC Medicine, 2025 (UK Biobank accelerometer, life-years), building on the foundational Moore et al., PLOS Medicine, 2012 (654,827 participants), one of the first sources to estimate life expectancy directly (~3.4 years gained at recommended activity levels, up to ~4.5 at high levels).
How we use it: A per-answer life-years adjustment from sedentary to active. The newer device-measured step data (~7,000 steps/day carries roughly half the mortality risk of ~2,000) is both stronger and more communicable than the older MET-hours framing.

Diet (Tier 1)

Source: Fadnes et al., PLOS Medicine, 2022 (the Food4HealthyLife modeling study), which reports life-expectancy gains from sustained dietary change. We use the conservative "feasibility" figures, scaled down with age. The same authors re-validated the model in Am J Clin Nutr, 2024, and it is empirically corroborated by a UK Biobank diet-index cohort (~104,000; +1.5–3 life-years for the top versus bottom quintile).
Note: We previously cited the PREDIMED trial here; we replaced it because PREDIMED measured cardiovascular events, not all-cause life expectancy. Fadnes reports the actual life-years.
How we use it: Rewards a shift toward whole foods, not adherence to any commercial diet.

Alcohol (Tier 1)

Source: GBD Alcohol Collaborators, The Lancet, 2018, with the 2022 age-specific update, and light-drinking neutrality anchored to a 2023 bias-corrected meta-analysis (Zhao/Naimi, ~4.8M) plus Mendelian-randomization cohorts.
How we use it: Heavy drinking carries a clear penalty. Light-to-moderate drinking is treated as roughly neutral: we do not encode "any amount is harmful," because the evidence for monotonic harm from the first drink does not hold up, especially for adults over 40.

Body weight (BMI) (Tier 1)

Source: Bhaskaran et al., Lancet Diabetes & Endocrinology, 2018 (3.6M never-smokers), corroborated by the Prospective Studies Collaboration (2009) and the Global BMI Mortality Collaboration (2016). Lowest mortality at BMI ~22.5–25; years-of-life-lost rise across obesity classes. This remains the strongest source; an honest note, though, is that recent cohorts show the low-overweight band (BMI 25–27.5) carries little excess mortality, and the nadir rises with age.
How we use it: A J-shaped curve, down-weighted by about half, because body weight's mortality effect overlaps heavily with diet and activity (which we already score). Counting it at full strength would double-count the same biology. Adults under 20 use pediatric percentiles and we suppress adult BMI categories entirely for that group.

Sleep (Tier 2)

Source: Yin et al., JAHA, 2017 (J-shaped curve, lowest risk ~7h), with life-years modeled in newer cohort work (Li et al., QJM, 2024) and the ~7h nadir confirmed by a GeroScience 2025 meta-analysis (79 cohorts).
How we use it: A small adjustment; both short and long sleep raise risk, with the optimum around 7 hours. An emerging finding is that sleep regularity may predict mortality better than duration (Windred et al., SLEEP, 2024), which we're watching but don't yet score.

Social connection (Tier 2)

Source: Wang et al., Nature Human Behaviour, 2023 (90 prospective cohorts, 2.2M; isolation ~32% higher mortality), reinforced by the WHO Commission on Social Connection (2025), building on the foundational Holt-Lunstad et al., Perspectives on Psychological Science, 2015 (3.4M participants; 26–32% higher mortality with isolation/loneliness).
How we use it: A small, flagged adjustment: the study reports mortality odds, so the life-years figure is an estimate. Felt connection is what matters, not relationship type.

Stress (Tier 2)

Source: Lee & Singh, Annals of Epidemiology, 2021 (US NHIS-NDI linkage, 513,081; a large life-expectancy gap by distress severity) plus Hockey et al., 2021 (dose-response), supported by the foundational Russ et al., BMJ, 2012 (68,222 adults; dose-response with psychological distress).
How we use it: A small, flagged adjustment. The Lee & Singh linkage lets us cite a real cohort's own life-expectancy translation rather than converting hazard ratios ourselves.

Sense of purpose (Tier 2)

Source: Alimujiang et al., JAMA Network Open, 2019 (6,985 adults; 15.2% lower mortality with strong life purpose).
Honest caveat: This is our weakest-evidenced factor. Recent reverse-causation analyses (2022–24) suggest that declining health lowers one's sense of purpose, so the true causal effect is likely smaller than the raw association. We deliberately keep this adjustment small for that reason.
How we use it: The smallest, most-flagged adjustment, and we keep it deliberately, because a sense of direction is both linked to longevity and central to why longevity is worth pursuing.

Preventive care (Tier 2)

What the evidence actually says: General annual checkups do not reduce all-cause mortality (Cochrane review, Krogsbøll et al.; confirmed by the NordICC screening trial, NEJM 2022, where all-cause mortality was nearly identical). The definitive source is Bretthauer et al., JAMA Internal Medicine, 2023 (18 RCTs, 2.1M), which found that cancer screening and checkups give almost no all-cause life-years gained (only sigmoidoscopy showed a small ~110-day gain). What does move all-cause mortality is managing specific risk factors: blood-pressure control (SPRINT) and statin primary-prevention trials both show a real signal for the risk-factor management we credit.
How we use it: We reframed this factor around knowing and managing your numbers (blood pressure, cholesterol, blood sugar) rather than checkup frequency, and we keep its adjustment small. We'd rather measure the thing that works than reward the thing that doesn't.

Family history (context only)

Source: Ruby et al., Genetics, 2018. Genetics accounts for roughly 7–12% of lifespan variation (less than the older "25%" claim).
How we use it: A small baseline adjustment based on reported grandparent longevity. It is not scored as a habit (you can't change it); it only nudges your baseline.

Putting the factors together: the interaction discount

We don't simply add the adjustments up. Habits overlap in their underlying biology: fixing sleep and stress doesn't give you the full sum of both, because they share pathways (cortisol, inflammation, metabolic health). Adding individual effects naïvely overstates the combined effect.

So we apply a single, moderate sub-additive discount to the positive adjustments, calibrated against the landmark combined-factor study: Li et al., Circulation, 2018, which found that the spread between a fully optimized profile and a highest-risk one is about +12 years (men) to +14 years (women) at age 50 for five low-risk habits versus none. This ceiling was corroborated in 2024 by the Million Veteran Program (Nguyen et al., Am J Clin Nutr, 2024; ~719,000, far more demographically diverse than Li's health-professional cohort), which found an even larger best-versus-worst gain, so our calibration is, if anything, conservative for younger users. We keep Li as the master anchor and do not re-anchor to the larger figure. Most of that spread comes from avoiding the downside of high-risk habits; the upside of optimizing from an already-average starting point is more modest, and we keep it deliberately conservative (under the Li ceiling) because we'd rather underpromise than overpromise. (A note: the +12/+14 figure is the Li-comparable spread; stacking genuinely extreme risks like severe obesity and daily smoking, which compound, can exceed it, which is expected and well-evidenced.)

A note on generalizability: the combined-factor study, like much of this literature, draws on cohorts that are predominantly white and higher-income. The effect sizes are the best available, but they may not transfer perfectly across every population: one more reason to read your number as directional, not exact.

How we build the habit library

Your estimate tells you where you stand. The habit library (160+ specific, science-backed actions across the nine habit areas) is what you can do about it. We hold those recommendations to the same standard as the number behind them.

Sourced from research, not vibes. Every action starts from the published literature (meta-analyses, randomized trials, and large prospective cohorts), not from generic wellness content. We deliberately go past the obvious basics ("sleep more") to the specific, non-obvious actions that show up in the research but rarely in listicles, so there's something new even if you already have the fundamentals down.

Every action is graded for evidence strength. Like the factors in your estimate, each action carries a label:

Core: backed by strong evidence such as randomized trials, meta-analyses, or large, replicated cohorts.
Emerging: promising but earlier, including smaller studies, mechanistic findings, or mixed results.

We show you which is which, so you can weight them yourself rather than taking everything as equally settled.

Citations are checked adversarially. It isn't enough to attach a plausible-sounding reference. Each action's source is interrogated against three questions: is the citation real, does it actually support the specific claim (rather than an exaggeration of it), and how strong is it? When the answer is uncertain, we grade the action down, or cut it entirely.

Every action passes a two-lens safety screen. This is the part we care about most. Before an action makes the library, it is reviewed for harm along two separate lenses, physical and mental, and anything that can't clear both is dropped or rewritten with the right guardrails. We screen for category-specific risks, not just generic ones: restriction or eating-disorder framing in diet; sleep anxiety ("orthosomnia") around sleep tracking; light-headedness or fainting in breathwork; shame or "willpower" framing around loneliness, addiction, and low mood; and dangerous withdrawal for heavy drinkers cutting back on alcohol. Where a real risk exists, the action carries a plain-language "before you start" note, and the sensitive areas point to genuine support: a clinician, or crisis lines like 988. When in doubt, we err toward caution.

No inflation, and not a substitute for your doctor. We'd rather under-promise: nothing in the library is hyped beyond what its evidence supports. The library is educational, a map for a conversation with a qualified professional who knows your situation, not medical advice or a treatment plan. Several actions are deliberately "ask your doctor about…" prompts for exactly that reason.

What Longlevity cannot tell you

Your actual lifespan. This is a statistical estimate anchored to population averages, not a personal forecast. We show a range, not just a single number, for exactly this reason.
Anything we didn't ask. The estimate is only as good as what you tell us: it's based on self-reported habits, not lab tests.
Your biological age in a clinical sense. We don't measure telomeres, epigenetic clocks, or any biomarker. Tools that do (TruDiagnostic, Elysium) are more precise on that axis; we don't compete there.
Disease-specific risk. We estimate all-cause mortality, not your risk for any particular condition. For cancer screening or cardiovascular risk scores, talk to your doctor.
Genetic variants or specific medical conditions. We use family history as a rough proxy and assume you're generally healthy. A diagnosed condition shapes your trajectory in ways we don't model.

We also deliberately exclude lab/biomarker inputs (VO2 max, ApoB, fasting glucose, grip strength). They're powerful but require testing, and they'd turn a 2-minute self-assessment into a clinical workup. That's a different product.

Editorial standards

No sponsored content. We don't accept payment to feature specific products, supplements, or services.

No affiliate placements in content. We may use affiliate links in the future (clearly disclosed), but editorial content is never influenced by affiliate relationships.

Primary sources only. Every claim traces to a specific published study. We don't cite other longevity blogs as sources.

Honest about uncertainty. Where research is contested (alcohol, preventive screening), we say so, and we use the most current and methodologically strongest evidence available, updating citations as the field evolves.

Updated on a schedule. Every content page is reviewed at least once every 12 months for accuracy. When research changes our model, we update both the math and this page so they never disagree.

Version history

We version the model so any estimate stays reproducible. Every saved result records the engine and calibration version that produced it, and older results recompute from their raw inputs when the model changes. This is the visible record of those changes.

Current: engine 2.2.0, calibration 2026.07.0 (July 2026). The quarterly citation refresh across eight factors. Alcohol light-drinking moved from a small positive to neutral (Zhao/Naimi 2023, ~4.8M, plus Mendelian-randomization cohorts). The sense-of-purpose adjustment was shrunk to reflect reverse-causation attenuation. Smoking, physical activity, diet, sleep, and social sources were modernized to recent cohorts (2023 to 2025). Alcohol, exercise, and sleep are now captured as continuous sliders (sleep as hours per night on a J-curve) rather than five-option buckets, with the locked deltas and endpoints preserved.

Notable changes since the first public model:

Diet source corrected. PREDIMED was replaced by Fadnes et al. (2022, re-validated 2024): PREDIMED measured cardiovascular events, not all-cause life-years, so it was the wrong anchor for a lifespan number.
Preventive care reframed. After the Cochrane review and the 2023 Bretthauer meta-analysis (18 RCTs, 2.1M) showed that general checkups do not move all-cause mortality, this factor was rebuilt around knowing and managing your numbers (blood pressure, cholesterol, blood sugar) instead of checkup frequency, with its weight kept deliberately small.
Baseline upgrades. A cohort adjustment (anchored to the Social Security Administration's intermediate assumptions, about +2 years at age 30 and fading to near zero by the 80s) so younger users are not systematically under-estimated, plus conditional survivorship life expectancy so remaining years are not underestimated for older users. Both are surfaced honestly, never as inflation.

The centering rebuild (our first public model). The foundational design: adjustments measured relative to the average person (average is the zero point, with no hidden penalty), a two-tier evidence grading (Tier 1 life-years versus Tier 2 flagged estimates), and a single sub-additive interaction discount anchored to the Li 2018 ceiling of about +12 to +14 years best-versus-worst.

Disclaimers

Longlevity is not a medical device. It's an educational tool. The estimates are based on statistical population-level research and do not account for your individual genetics, medical history, medications, or pre-existing conditions.

You should not rely on Longlevity for medical decisions. Always consult a qualified healthcare provider before making changes to your diet, exercise, medications, or health routine. If you're experiencing a medical emergency, call your local emergency number immediately.

Longlevity does not create a doctor-patient relationship. We're not HIPAA-covered and make no claims of HIPAA compliance.

Contact

If you have questions about our methodology, spot an error, or want to discuss a specific study we use, email contact@longlevity.ai.